Skip to content

Integration of deep generative Anomaly Detection algorithm in high-speed industrial line

Conference: CVPR2025
arXiv: 2603.07577
Code: None (Industrial closed-source)
Area: Others
Keywords: Anomaly Detection, GAN, Residual Autoencoder, Industrial Deployment, BFS Production Line, Real-time Inference

TL;DR

A semi-supervised anomaly detection framework based on GAN and Dual-stage Residual Autoencoder (DRAE) deployed for real-time online quality inspection on high-speed pharmaceutical BFS lines. Trained exclusively on normal samples, it achieves a single-patch inference time of 0.17 ms, optimizing reconstruction quality via Perlin noise augmentation and Noise Loss.

Background & Motivation

Background: BFS (Blow-Fill-Seal) lines in the pharmaceutical industry require non-destructive visual quality inspection. Currently, they still heavily rely on manual visual inspection, which is limited by fluctuations in operator attention and throughput bottlenecks.

Limitations of Prior Work: Classic rule-based computer vision algorithms (thresholding + template matching) are highly dependent on specific product configurations, feature excessive parameters, scale poorly, and struggle to distinguish process noise (e.g., bubbles in liquid) from genuine defects (e.g., foreign particles).

Limitations of Supervised Learning: In industrial scenarios, conforming products vastly outnumber defective ones, leading to severe class imbalance where supervised classification methods cannot be directly applied.

Key Challenge: High-precision anomaly detection and spatial localization must be achieved under strict hardware constraints (Industrial PC + A4500 GPU) and time constraints (500 ms acquisition interval).

Key Insight: A semi-supervised approach is adopted where a GAN is trained solely on normal samples to detect anomalies via reconstruction residuals, with engineering optimization tailored for industrial production lines to achieve a truly deployable system.

Method

Overall Architecture

The system improves upon GRD-Net and consists of a GAN network: - Generator G: An Encoder-Decoder-Encoder structure, at the core of which is a Dual-stage Residual Autoencoder (DRAE) using a fully-connected bottleneck layer (64-dimensional features). The encoder uses a 4-stage ResNet v2 architecture (outputting 16×16×1024), and the decoder is a symmetric transposed convolutional structure. - Discriminator C: A convolutional encoder + fully connected layer, performing true/false binary classification.

Loss & Training

  1. Adversarial Loss \(\mathcal{L}_{adv}\): Matches feature activations in the final convolutional layer of the discriminator.
  2. Context Loss \(\mathcal{L}_{con}\): A weighted combination of Huber Loss (replacing the original L1 for improved stability) and SSIM Loss.
  3. Encoder Consistency Loss \(\mathcal{L}_{enc}\): Restricts the latent representations of the original and reconstructed images to be consistent.
  4. Noise Loss \(\mathcal{L}_{nse}\) (newly introduced in this work): Explicitly supervises the network's reconstruction behavior on Perlin noise regions, preventing entropy increase from the augmentation from destabilizing optimization.

Total generator objective: \(\mathcal{L}_{gen} = w_1 \cdot \mathcal{L}_{adv} + w_2 \cdot \mathcal{L}_{con} + w_3 \cdot \mathcal{L}_{enc} + w_4 \cdot \mathcal{L}_{nse}\)

Perlin Noise Augmentation

  • Overlays Perlin noise on normal samples with a probability of \(q=0.75\) to formulate a denoising task.
  • The noise blending factor \(\beta \sim \mathcal{U}(0.5, 1.0)\) controls noise intensity.
  • Additional random rotations in \([-\pi/8, \pi/8]\) and vertical flips are applied, excluding horizontal flips that could generate unrealistic patterns.

Anomaly Scoring and Localization

  • Anomaly score: \(\phi = 1 - \text{SSIM}(X, \hat{X})\)
  • Heatmap: min-max normalization of \(H = |X - \hat{X}|\)
  • Thresholds are tuned on an independent calibration set and configured differently for each region.

Preprocessing Pipeline

  • Each BFS strip contains 5 ampoules, with each ampoule further divided into 4 logical regions (label region, upper body, liquid level region, bottom).
  • 20 gray-scale patches of size 256×256 are extracted from each frame.
  • Training set: 782 strips × 10 acquisitions × 16 frames × 20 patches = 2,815,200 images.

Key Experimental Results

Detection Performance (Real Industrial Test Set: 141 Defective + 120 Conforming)

Level Accuracy TPR TNR Balanced Accuracy
Patch (Overall) 99.19%~99.91% 99.66%~99.94% 90.44%~99.73% 95.15%~99.84%
Product Level 95.93% 96.94% 94.67% 95.81%
Run Level (7/10 Voting) 96.41% 96.76% 95.99% 96.38%

Inference Speed

  • Single-patch inference time: 0.1689 ms (A4500 GPU)
  • Single-product inference time: 0.4873 ms (<500 ms acquisition interval, satisfying real-time constraints)

Training Details

  • Training set of 2.815 million patches, trained for 10 epochs.
  • Learning rate of 1.5×10⁻⁴ with cosine decay restarts, batch size of 32.
  • Training hardware: A100 40GB; Inference hardware: A4500 20GB.

Industrial Deployment Details

  • The inference side is integrated into the machine control software via the C++ TensorFlow API.
  • A rotary online quality inspector grabs products from the conveyor belt into a carousel, performing automated sorting post-detection.
  • Telecentric lenses are utilized to eliminate geometric distortion, particularly at the sides and bottom.
  • A rank filter is applied during training to generate the brightest/darkest images, increasing the variability of normal samples.
  • A run-level voting strategy (at least 7 out of 10 consistent acquisitions) is designed to simulate the decision logic of human inspectors.
  • Compared to embedding similarity methods like PaDiM/PatchCore: This work opts for a reconstruction-based approach because it provides intuitive, explainable heatmaps, matching the industrial customers' requirement for visual feedback on anomaly locations.
  • Compared to DRÆM: Although sharing the Perlin noise augmentation concept, DRÆM relies on an extra U-Net segmentation network. This work removes the segmentation branch and retains only the generator to meet inference time constraints.
  • Compared to EfficientAD: While EfficientAD aims for lightweight execution, it is still based on embedding distances. This work provides pixel-level explanations based on reconstruction residuals.

Highlights & Insights

  • Genuine Industrial Deployment: Rather than a theoretical laboratory paper, this work details the entire workflow from data acquisition and model training to production line integration (C++ TensorFlow API), including GMP compliance validation.
  • Practical Innovation in Noise Loss: By explicitly constraining reconstruction behavior in noisy regions, it resolves the training instability issue introduced by Perlin noise augmentation.
  • Awareness of Engineering Constraints: Design trade-offs are made under strict hardware (Industrial PC + A4500) and time (500 ms slot) limitations, compressing the model significantly using a fully connected bottleneck layer (64-dimensional).
  • Multi-Level Evaluation Protocol: A three-level aggregation protocol (patch \(\rightarrow\) product \(\rightarrow\) run level) is established, adhering to industrial acceptance standards.
  • Heatmap Explainability: Anomaly region heatmaps are automatically generated via reconstruction discrepancies, providing operators with intuitive defect localization on the HMI.

Limitations & Future Work

  • Validated only on a single product (BFS vials), without comparative evaluation on public benchmarks such as MVTec.
  • Restricted by NDA, the dataset and full pipeline cannot be publicly released, making verification or follow-up difficult for academia.
  • Quantitative comparison with current SOTA methods like PatchCore or EfficientAD is absent, failing to evaluate competitive performance under general scenarios.
  • TNR is relatively low in certain regions (R0: 90.93%, R1: 90.44%), where an approximately 10% false positive rate could cause a massive amount of false rejections on high-speed lines.
  • Threshold tuning relies on manual configuration on a validation/calibration set, lacking an adaptive mechanism to handle batch-to-batch product variations.
  • The liquid level region exhibits high variability due to bubble movement; the paper does not thoroughly discuss the robustness in this region.
  • Employs TensorFlow 2.x instead of PyTorch, limiting ecosystem compatibility.

Rating

  • Novelty: ⭐⭐⭐⭐ The method itself is an engineering improvement on GRD-Net/DRÆM, though Noise Loss exhibits certain novelty.
  • Experimental Thoroughness: ⭐⭐⭐⭐ Evaluation on real-world industrial data is compelling, but public benchmark comparisons are lacking.
  • Writing Quality: ⭐⭐⭐⭐ Engineering details are abundant, but mathematical notation is occasionally disorganized, and the structure is slightly redundant.
  • Value: ⭐⭐⭐⭐⭐ Highly valuable reference for industrial anomaly detection deployment, bridging the gap between academic research and production line deployment.